htcondor and workflows: advanced tutorial htcondor week …...13 rescue dags (cont) › save the...

HTCondor and Workflows:Advanced Tutorial

HTCondor Week 2014Kent Wenger

Workflows in HTCondor

●This talk: techniques & features

●Please ask questions!

Example workflow

...10k...

Preparation

Simulation

Analysis

NOOP?

CleanupFinal

4

How big?› We have users running

500k-job workflows in production (user hit 2.2 GB DAG file bug)

› Depends on resources on submit machine (memory, max. open files)

› “Tricks” can decrease resource requirements (talk to me)

Pretty big!

5

Defining a DAG to DAGMan

A DAG input file defines a DAG:

# file name: diamond.dagJob A a.submitJob B b.submitJob C c.submitJob D d.submitParent A Child B CParent B C Child D

A

B C

D

Organization of files and directories

Files and directories

●By default, all paths in a DAG input file and the associated submit files are relative to the current working directory when condor_submit_dag is run.

●Modified by DIR directive on JOB command●Also by -usedagdir on condor_submit_dag command line

Nodes in subdirectories

NodeAA.subA.in1A.preA.post...

NodeBB.subB.in1B.preB.post...

Topmy.dag

# my.dag:Job A A.sub Dir NodeAScript Pre A A.preJob B B.sub Dir NodeB...# A.sub...input = A.in1...

# in Top:condor_submit_dag my.dag

DAGs in subdirectories

Dag1my.dagA.subB.subA.in...

Top

# Dag1/my.dag:Job A A.subJob B B.sub...

# in Top:condor_submit_dag -usedagdir Dag1/my.dag Dag2/my.dag ...

Dag2my.dagA.subB.subA.in...

Multiple independent jobs

●Why use DAGMan?:● Throttling● Retry of failed jobs● Rescue DAG● PRE/POST scripts● Submit file re-use

●DAG file has JOB commands but not PARENT...CHILD

Rescue DAGs

12

Rescue DAGs (cont)

Run

Not run

A

B1

D

B2 B3

C1 C2 C3

13

Rescue DAGs (cont)› Save the state of a partially-completed DAG› Created when a node fails or the condor_dagman job is removed with condor_rm or when DAG is halted and all queued jobs finishDAGMan makes as much progress as possible in the

face of failed nodes› DAGMan immediately exits after writing a rescue

DAG file› Automatically run when you re-run the original

DAG (unless –f is passed to condor_submit_dag)

Rescue DAGs (cont)› The Rescue DAG file, by default, is only a partial

DAG file.› A partial Rescue DAG file contains only

information about which nodes are done, and the number of retries remaining for nodes with retries.

› Does not contain information such as the actual DAG structure and the specification of the submit file for each node job.

› Partial Rescue DAGs are automatically parsed in combination with the original DAG file, which contains information such as the DAG structure.

14

Rescue DAGs (cont)

› If you change something in the original DAG file, such as changing the submit file for a node job, that change will take effect when running a partial rescue DAG.

15

16

Rescue DAG naming

› DagFile.rescue001, DagFile.rescue002, etc.

› Up to 100 by default (last is overwritten once you hit the limit)

› Newest is run automatically when you re-submit the original DagFile

› condor_submit_dag -dorescuefrom number to run specific rescue DAGNewer rescue DAGs are renamed

DAGMan configuration

18

DAGMan configuration (cont)

› A few dozen DAGMan-specific configuration macros (see the manual…)

› From lowest to highest precedenceHTCondor configuration filesUser’s environment variables:

• _CONDOR_macronameDAG-specific configuration file (preferable)condor_submit_dag command line

19

Per-DAG configuration› In DAG input file:CONFIG ConfigFileName or command line:condor_submit_dag –config ConfigFileName ...

› Generally prefer CONFIG in DAG file over condor_submit_dag -config or individual arguments

› Specifying more than one configuration file is an error.

Per-DAG configuration (cont)

› Configuration entries not related to DAGMan are ignored

› Syntax like any other HTCondor config file# file name: bar.dagCONFIG bar.config

# file name: bar.configDAGMAN_ALWAYS_RUN_POST = FalseDAGMAN_MAX_SUBMIT_ATTEMPTS = 2

20

Configuration: workflow log file

●DAGMan now uses a single log file for all node jobs

●Put workflow log file on local disk●DAGMAN_DEFAULT_NODE_LOG●In 8.1(/8.2): changes for global config

● @(DAG_DIR)/@(DAG_FILE).nodes.log● /localdisk/@(DAG_FILE).nodes.log

Pre skip

DAG node with scripts: PRE_SKIP

› Allows PRE script to immediately declare node successful (job and POST script are not run)

› In the DAG input file:JOB A A.cmdSCRIPT PRE A A.prePRE_SKIP A non-zero_integer

› If the PRE script of A exits with the indicated value, the node succeeds immediately, and the node job and POST script are skipped.

› If the PRE script fails with a different value, the node job is skipped, and the POST script runs (as if PRE_SKIP were not specified).

23

DAG node with scripts: PRE_SKIP (cont)

› When the POST script runs, the $PRE_SCRIPT_RETURN variable contains the return value from the PRE script. (See manual for specific cases)

24

No-op nodes

No-op nodes (cont)

› Appending the keyword NOOP causes a job to not be run, without affecting the DAG structure.

› The PRE and POST scripts of NOOP nodes will be run. If this is not desired, comment them out.

› Can be used to test DAG structure

26

No-op nodes (ex)

› Here is an example:# file name: diamond.dagJob A a.submit NOOPJob B b.submit NOOPJob C c.submit NOOPJob D d.submit NOOPParent A Child B CParent B C Child D

› Submitting this to DAGMan will cause DAGMan to exercise the DAG, without actually running node jobs.

27

No-op nodes (ex 2)Simplify dag structure

NOOP

29

Node retries

› For possibly transient errors› Before a node is marked as failed. . .

Retry N times. In the DAG file:Retry C 4(to retry node C four times before calling the node

failed)Retry N times, unless a node returns specific exit

code. In the DAG file:Retry C 4 UNLESS-EXIT 2

30

Node retries, continued› Node is retried as a whole

Job

PRE

POST

Node

SuccessUnless-exit value:

node fails

One node failure:retry

Out of retries:node fails

Node variables

32

Node variables (cont)› To re-use submit files› In DAG input file:VARS JobName varname="value" [varname="value"... ]

› In submit description file:$(varname)

› varname can only contain alphanumeric characters and underscore

› varname cannot begin with “queue”› varname is not case-sensitive› varname beginning with “+” defines classad attribute

(e.g., +State = “Wisconsin”)

Node variables (cont)

› Value cannot contain single quotes; double quotes must be escaped

› The variable $(JOB)contains the DAG node name

› $(RETRY) contains retry count› Any number of VARS values per node› DAGMan warns if a VAR name is defined

more than once for a node

33

Node variables (ex)

# foo.dagJob B10 B.subVars B10 infile=”B_in.10”Vars B10 +myattr=”4321”

# B.subinput = $(infile)arguments = $$([myattr])

Nested DAGs

36

Nested DAGs (cont)

37

Nested DAGs (cont)› Runs the sub-DAG as a job within the top-level

DAG› In the DAG input file:SUBDAG EXTERNAL JobName DagFileName

› Any number of levels› Sub-DAG nodes are like any other (can have

PRE/POST scripts, retries, DIR, etc.)› Each sub-DAG has its own DAGMan

Separate throttles for each sub-DAGSeparate rescue DAGs

Why nested DAGs?

› DAG re-use› Scalability› Re-try more than one node› Short-circuit parts of the workflow› Dynamic workflow modification (sub-DAGs

can be created “on the fly”)

38

Splices

40

Splices (cont)

› Directly includes splice DAG’s nodes within the top-level DAG

› In the DAG input file:SPLICE JobName DagFileName

› Splices can be nested (and combined with sub-DAGs)

Why splices?●DAG re-use●Advantages of splices over sub-DAGs:

● Reduced overhead (single DAGMan instance)● Simplicity (e.g., single rescue DAG)● Throttles apply across entire workflow

●Limitations of splices:● Splices cannot have PRE and POST scripts (for

now)● No retries● Splice DAGs must exist at submit time

Throttling

43

Throttling (cont)› Limit load on submit machine and pool

Maxjobs limits jobs in queue

Maxidle submit jobs until idle limit is hit• Can get more idle jobs if jobs are evicted

Maxpre limits PRE scriptsMaxpost limits POST scripts

› All limits are per DAGMan, not global for the pool or submit machine

› Limits can be specified as arguments to condor_submit_dag or in configuration

Node categories

45

Node categories (cont)

Setup

Cleanup

Big job

Small jobSmall jobSmall job

Big job

Small jobSmall jobSmall job

Big job

Small jobSmall jobSmall job

46

Node category throttles

› Useful with different types of jobs that cause different loads

› In the DAG input file:CATEGORY JobName CategoryNameMAXJOBS CategoryName MaxJobsValue

› Applies the MaxJobsValue setting to only jobs assigned to the given category

› Global throttles still apply

47

Cross-splice node categories

› Prefix category name with “+”MaxJobs +init 2Category A +init

› See the Splice section in the manual for details

Node priorities

49

Node priorities (cont)

› In the DAG input file:PRIORITY JobName PriorityValue

› Determines order of submission of ready nodes

› DAG node priorities are copied to job priorities (including sub-DAGs)

› Does not violate or change DAG semantics› Higher numerical value equals “better”

priority

Node priorities (cont)

› Better priority nodes are not guaranteed to run first!

› Effective node prio = max(explicit node prio, parents' effective prios, DAG prio)

› For sub-DAGs, pretend that the sub-DAG is spliced in.

› Overrides priority in node job submit file

50

Node priorities (upcoming changes)

●Priority change to DAGMan job “trickles down” to nodes

●Different “inheritance” policy:● Effective node prio = explicit node prio +

DAG prio?● Effective node prio = average(explicit node

prio, parents' effective prios, DAG prio)?

52

DAG abort

●In DAG input file:ABORT-DAG-ON JobName AbortExitValue [RETURN DagReturnValue]

●If node value is AbortExitValue, the entire DAG is aborted, implying that queued node jobs are removed, and a rescue DAG is created.

●Can be used for conditionally skipping nodes (especially with sub-DAGs)

FINAL nodes

FINAL nodes (cont)

› FINAL node always runs at end of DAG (even on failure)

› Use FINAL in place of JOB in DAG file› At most one FINAL node per DAG› FINAL nodes cannot have parents or

children (but can have PRE/POST scripts)

54

FINAL nodes (cont)

› Success or failure of the FINAL node determines the success of the entire DAG

› PRE and POST scripts of FINAL (and other) nodes can use $DAG_STATUS and $FAILED_COUNT to determine the state of the workflow

› $(DAG_STATUS) and $(FAILED_COUNT) in available in VARS

55

Advanced workflow monitoring

Status in DAGMan’s ClassAd> condor_q -l 59 | grep DAG_DAG_Status = 0DAG_InRecovery = 0DAG_NodesUnready = 1DAG_NodesReady = 4DAG_NodesPrerun = 2DAG_NodesQueued = 1DAG_NodesPostrun = 1DAG_NodesDone = 3DAG_NodesFailed = 0DAG_NodesTotal = 12

› Sub-DAGs count as one node› New in 7.9.5

57

58

Node status file

› Shows a snapshot of workflow stateOverwritten as the workflow runsUpdated atomically

› In the DAG input file:NODE_STATUS_FILE statusFileName [minimumUpdateTime]

› Not enabled by default› As of 8.1.6, in ClassAd format (a set of ClassAds)

Node status file contents[ Type = "DagStatus"; DagFiles = { "job_dagman_node_status.dag" }; Timestamp = 1397683160; /* "Wed Apr 16 16:19:20

2014" */ DagStatus = 3; /* "STATUS_SUBMITTED ()" */ NodesTotal = 12; NodesDone = 0; NodesPre = 0; NodesQueued = 1; ...

Node status file contents (cont)[ Type = "NodeStatus"; Node = "C"; NodeStatus = 6; /* "STATUS_ERROR" */ StatusDetails = "Job proc (1980.0.0) failed with status 5";

RetryCount = 2; JobProcsQueued = 0; JobProcsHeld = 0;]

61

Jobstate.log file

› Shows workflow history› Meant to be machine-readable (for Pegasus)› Basically a subset of the dagman.out file› In the DAG input file:JOBSTATE_LOG JobstateLogFileName

› Not enabled by default

62

Jobstate.log contents

1302884424 INTERNAL *** DAGMAN_STARTED 48.0 ***

1302884436 NodeA PRE_SCRIPT_STARTED - local - 1

1302884436 NodeA PRE_SCRIPT_SUCCESS - local - 1

1302884438 NodeA SUBMIT 49.0 local - 11302884438 NodeA SUBMIT 49.1 local - 11302884438 NodeA EXECUTE 49.0 local - 11302884438 NodeA EXECUTE 49.1 local – 1...

DAGMan metrics

●Anonymous workflow metrics (for Pegasus)

●Metrics file (JSON format) generated at end of run (dagfile.metrics)

●Reported by default (can be disabled)

●Dagman.out tells whether metrics were reported

DAGMan metrics example{ "client":"condor_dagman", "version":"8.1.6", ... "start_time":1396448008.138, "end_time":1396448047.596, "duration":39.458, "exitcode":0, ... "total_jobs":3, "total_jobs_run":3, "total_job_time":0.000, "dag_status":0}

65

More information

› There’s much more detail, as well as examples, in the DAGMan section of the online HTCondor manual.

› DAGMan: http://research.cs.wisc.edu/htcondor/dagman/dagman.html

› For more questions: htcondor-admin@cs.wisc.edu, htcondor-users@cs.wisc.edu

Extra slides

DAGMAN_HOLD_CLAIM_TIME› An optimization introduced in HTCondor

version 7.7.5 as a configuration option› If a DAGMan job has child nodes, it will

instruct the HTCondor schedd to hold the machine claim for the integer number of seconds that is the value of this option, which defaults to 20.

› Next job starts w/o negotiation cycle, using existing claim on startd

67

67

Dot file

› Shows a snapshot of workflow state› Updated atomically› For input to the dot visualization tool› In the DAG input file:DOT DotFile [UPDATE] [DONT-OVERWRITE]

› To create an imagedot -Tps DotFile -o PostScriptFile

69

Dot file example